A production brain's first process boot after a live authority flip looked
hung and was restarted three times mid-recovery — three defects with one
scene. (1) THE FOLD MATERIALIZED THE LOG: peekFactsAbove(0) decoded every
fact into one array (GBs of after-images on a ~7k-fact log, a GC storm, a
starved write lane). The fold now STREAMS one segment-batch at a time —
memory is one segment at any log size — with structural ordering asserted
loudly. (2) THE FOLD WAS SILENT UNTIL DONE: minutes of boot work with zero
narration is what invited the restarts. It now announces itself BEFORE the
work ('do not restart, the fold is finite') and prints progress every
thousand facts. (3) THE CHAIN COULD ONLY ARM AT A CRASH: a live mid-session
flip left the fold checkpoint unfounded, so the brain's first unclean boot
paid a whole-log fold. Adoption now founds the checkpoint AT THE FLIP — one
paged full canonical barrier (bounded memory), then the stamp — so bounded
recovery holds from minute zero for every store that flips, at any size.
Pinned: a non-fresh flip stamps immediately; the first post-flip unclean
boot folds bounded (an unflushed at-ack fact above the checkpoint is
restored; a barrier-covered row below it is outside the fold). Kill matrix
and both adoption suites green alongside.
The plant's integration lane caught it twice: the fence's startedAt-strict
comparison turned the documented same-process warn-and-take-over path (two
instances in one Node process — the server-restart test pattern, and the
shared-default-store pattern across test files) into a flush-killer: the
first instance's background flushes latched dead while its own process held
the lock ('PID N no longer holds the lock — it is now held by PID N').
Ownership is per-process: pid + hostname. startedAt stays in the lock for
observability but not in the fence — it protects nothing (a pid-recycled
successor's victim is a dead process that runs no fence checks) and it
convicted the innocent. Pinned: a same-process re-open leaves both
instances' flushes working; the cross-process eviction pins unchanged.
Verified under the lane's exact command: 102/102 files, 850 passed, exit 0.
Seven micro-budget tests were calibrated on one fast desktop and failed on
other honest iron with zero functional failures (bisect-proven pre-existing;
David-waived for 10.1/10.2 with this recalibration filed as the cure). Every
budget is now at least 3x the worst measurement observed across three
machines, each with a comment naming its calibration basis; the find-unified
micro-comparison of two sub-millisecond timings becomes a ratio assertion
(absolute equality of microsecond pairs can never be stable). The
inference-bound trim-history correctness test gets a timeout covering its
slowest observed run (174s) — its assertions are exact and untouched.
These remain order-of-magnitude guards; real perf enforcement lives in the
dedicated perf lanes with iron-specific budgets, per the gate-speed standard.
Known non-test artifact, documented not hidden: on slow-inference machines a
minutes-long awaited-embed loop can trip vitest's worker-RPC 60s tolerance
('Timeout calling onTaskUpdate') — all tests pass, vitest exits 1 on the
unhandled orchestration error. The CI lanes on faster iron exit clean; if a
lane ever trips it, the test moves to deterministic embeddings (its
assertions are size-bookkeeping, not embedding quality).
The production dev-store split-brain (two live writers alternating a store's
id-mapper between two internally-consistent truths), cured at all three of
its roots. (1) STALENESS REQUIRES PID-DEATH: the old rule evicted on
heartbeat age alone, so a >60s event-loop stall (debugger pause, GC, heavy
sync work) handed the lock to a second opener while the first kept writing;
a live process is now never auto-evicted — a wedged-but-alive holder is the
operator's call via {force:true}, and the heartbeat stays for observability.
(2) THE CLAIM IS ATOMIC: writeFile(wx)'s open→write→close left an empty-file
window a concurrent opener could read as torn, unlink a LIVE claim, and take
the lock; the claim is now tmp-write + hard-link — the lock appears with its
full contents in one step. (3) THE FENCE: every flush commit and transact
barrier verifies lock ownership first (one small read per window) — a
forced-out or lock-deleted writer fails typed (BRAINY_WRITER_FENCED) before
a single staged byte or manifest advance, instead of writing on unaware.
Pinned: live-with-ancient-heartbeat refuses typed; dead-PID self-clears
narrated; a forced-out writer's flush and transact both fence, advancing
nothing. Requested by a downstream team as single-writer guard or loud
lockout — this is both.
Two consumer-driven cures sharing one stamp. (1) TX-LOG ORIGIN: engine-
originated commits stamp an optional origin on their tx-log entry AND the
commit fact's meta — 'system:embed-landing' (the deferred vector landing),
'system:adoption-backfill' (baseline re-commits), 'system:reconcile'. A
downstream activity feed showed a double tick because the landing commit was
indistinguishable from a user save, and the consumer rightly refused a
time-window collapse as a quiet loss; feeds now filter on fact. User writes
stay unstamped — absent origin is the user shape, every existing consumer
unchanged. (2) reconcileLogDivergence(id, {attest}): the human's door for
log-live-canonical-absent, the one class adoption refuses by design because
a lost-tombstone deletion is indistinguishable from canonical loss.
'deleted' mints the missing tombstone (history keeps the earlier live
record); 'restore' folds the log's only copy back into canonical; wrong-
class calls refuse typed with nothing written. Loud, narrated, single-row,
origin-stamped. From a production adoption's one surviving divergence.
A production brain with a 12.7k-row pre-log baseline advanced exactly 800
rows per adoptLogAuthority() call (a five-pass ceiling × the oracle's
200-row listing cap), refused the flip, and sat tree-authoritative for
hours across restarts. The bound was sized for drift, never for a baseline.
Now: the adoption path runs the oracle uncapped so ONE scan yields the
ENTIRE curable set, every pass cures all of it, and the loop runs to
completion with the no-progress guard as its only stop. Pace rides the
write path (one full-brain scan amortizes over thousands of cures, not two
hundred): 1,000 drifted rows adopt green in one call in ~10s. Progress is
narrated for a live operator. The wire report keeps its 200-row cap.
Pinned: a baseline above the old ceiling adopts green in a single call.
Two defects with one root, found by the fold-checkpoint work's first
integration gate. (1) THE RACE: restore() never quiesced the generation
store, so a background flush could write into _system/ while the swap was
removing it — observed as ENOTEMPTY mid-swap when a checkpoint stamp landed
between readdir and rmdir. The swap now runs inside the store's exclusive
section (runStateReplacement): flush timer disarmed, pending tier and
checkpoint accumulator discarded BEFORE any directory moves. (2) THE
INHERITED ASSERTION: a snapshot carries its source brain's clean-shutdown
marker and fold checkpoint, but the restored files were bulk-copied without
per-file fsync — the inherited stamps would suppress exactly the recovery
fold that cures a post-restore power cut. reopenAfterRestore now deletes
both stamps before reopening: the open treats the store as uncleanly shut,
folds the restored log into canonical, barrier-syncs what it re-applied,
and stamps fresh — the restored state is durably founded at restore time
instead of borrowing assertions about bytes this disk never synced.
Pinned: restore under in-flight traffic completes; the pre-restore stamp
does not survive; the post-restore stamp is the reopen fold's own, at the
restored watermark.
The fold checkpoint (_system/fold-checkpoint.json) is stamped strictly after
a canonical-sync barrier over every live entity touched since the last stamp
(syncEntityCanonical: ids → canonical paths → fsync; an absent file fsyncs
its parent directory so deletes are as durable as writes). An unclean open
under log authority now folds only (checkpoint, head]; the chain bootstraps
at an empty brain's adoption (three-phase hooks around adoptLogAuthority) or
at a brain's first whole-log fold — existing brains converge at their first
crash with zero regression. Rollback restores sync immediately; abort paths
feed the barrier; a failed barrier retains the old bound (bigger fold later,
never a lost write). Five structural pins including boundedness itself.
Also: the production-shaped write-flow gate leg (mixed traffic racing
flushes, crash mid-traffic, every ack survives — from a consumer-reported
gate miss), and two release-ceremony cures (tag-first push so the publish
never queues behind the release commit's CI run; raw-curl npmjs shasum
probe with propagation grace instead of a one-shot false divergence).
An adopter's full suite found two v2 write-path defects on fresh brains,
reproduced with stacks; both cured and both pinned with their exact
production shapes:
1. PAD-FRAME CONSTRUCTIBILITY: a single msgpack bin filler steps its
header by one byte at each size class (bin8→bin16→bin32), leaving one
unreachable payload size per boundary — the sealer requested a
291-byte pad, the encoder threw 'not constructible', and sync() died
whole. Construction is now TOTAL: the class-boundary holes bridge with
a trailing fixint beside the bin ({bin(n)} ∪ {bin(n)+fixint} covers
every size ≥ minimum). Pinned exhaustively: every size from the
minimum through a full sector plus boundary spill constructs
byte-exact and decodes as reader-invisible filler.
2. THE NON-MONOTONIC REFUSAL LOOP: the append-failure compensation
rewound the generation counter on ANY throw — including a covering
SYNC failure after a SUCCESSFUL append. The log carried generation N
while the counter re-minted N, and every later append refused
'non-monotonic (N ≤ head N)' — the write path wedged in a refusal
loop through deferred-embed retries and flush backoff. The
compensation now splits by phase: an append failure (log never took
the fact) fully compensates — un-buffer and rewind; a sync failure
after append earns the rewind ONLY if the appended fact is provably
dropped, otherwise the generation stays consumed and buffered — the
counter never re-mints a number the log may carry. Pinned: an
injected one-shot sync failure fails its write loudly and the very
next write mints fresh and succeeds, with the log scanning strictly
ascending end to end.
Also probed against the adopter's carried report: the 9.0 vfs.rename
stale-ghost shape does NOT reproduce on this head (old path cleanly
unresolvable on exists/stat/readdir after rename).
Gates: unit 2067/2067 (160 files) · integration 833 (97 files) ·
conformance 36/36.
A first adopter's namespace migration went 341 red on one class: the
never-carried-field refusal firing on CORRECT filters against fresh and
sparse stores — a freshly provisioned tenant refused its own first
filtered read, with the did-you-mean built for typos firing hardest on
day-one stores where nothing is wrong.
The ruled cut: a WHERE filter naming a field no row carries is SERVED
OPERATOR-TRUTHFULLY — eq/in/range/contains answer [] (nothing carries
it, nothing matches); ne and exists:false answer ALL rows (the equally
true complement — a blanket empty here would be silently wrong, which is
why the simpler cut was rejected); exists:true answers []. Served from
the field registry, with the did-you-mean demoted to a once-per-field
WARN. orderBy and genuinely ambiguous addresses KEEP their hard typed
refusals: no truthful order exists over an uncarried field, and
ambiguity is a contract error while absence is data.
Mechanics: the negative operator absorbs the FIELD_NOT_INDEXED throw as
its empty exclude set (the clause-level catch correctly zeroes positive
operators only); the egress matcher already agreed. Plus the
provider-seam belt: a field refusal thrown by a replacement metadata
manager is normalized to THIS package's UnresolvableFieldError at every
filter call site — one class identity for consumers, instanceof works
(a first adopter's cross-package finding).
Conformance: tests/conformance/sparse-store-cut.test.ts — the shared
operator rows both engines run (positive-empty, negative-all,
fresh-tenant day-one, orderBy refusal kept, compound composition).
Gates: unit 2065/2065 · integration 832 · conformance 36/36.
The last rung of the default-flip ruling: with the sentinel exemption in,
real production-shaped brains still refused adoption over state-differs
mismatches the backfill could not cure — rows written before the
hydration law carry denormalized wrapper fields that disagree with their
own metadata leg, and the previous as-is identity re-commit PRESERVED
that drift, so the oracle re-flagged it every pass and the flip never
happened. In practice the crash-safe default reached zero existing
brains: the exact outcome the hold ruling forbade.
The cure: the backfill now rewrites canonical in the LAW SHAPE — exactly
the wrapper the log's reconstruction produces (denormalized enumeration
fields derived from the metadata leg, which is their authority under the
field-addressing law; the embedding floats ride through byte-identical;
adjacency residue keeps its own rebuild path). The oracle then verifies
the rewrite before the flip — the same safety, no operator chore.
Log-ahead divergence classes (a log the witness denies) still refuse
loudly, exactly as before.
Classification note for the record: the flagged uuid-v7 rows postdate the
fact log's introduction, so they classify as state-differs (in-log,
drift-shaped) rather than pre-log — both classes ride the same backfill.
Pins: a manufactured depot-shape drifted wrapper adopts green with floats
preserved and metadata intact; log-ahead still refuses typed.
Gates: unit 2065/2065 · integration 832 · conformance 31/31.
The release-holding finding from the joint gate's six real depot brains:
the adoption path's positive-int mint check false-flagged the reserved
VFS-root sentinel (the all-zeros UUID, minted int 0 BY CONSTRUCTION at
genesis on existing brains) as a corrupt mint — so every existing brain
refused log-authority adoption and stayed on the old lossy-under-power-cut
durability, defeating the release's headline crash-safety exactly where
it matters most.
The exemption, at both mint seams (the host's minter thunk and the fact
log's encoder guard): int 0 is legal iff the id is the reserved root;
zero for ANY other id remains a corrupt-mint refusal naming the reserved
exception. The codec's u64 layer already tolerated 0 — only the guards
over-refused.
Pins: adoption goes green on a brain whose VFS root carries int 0 (the
depot-brain shape, previously refused) · a non-root zero still refuses
typed at the mint seam — held at the seam itself because a full write
SELF-HEALS a poisoned zero (the index cycle re-mints before the fact is
written, which is the correct outcome and was verified in the pinning).
Gates: unit 2065/2065 · integration 830 · conformance 31/31.
The quiet-loss cure regressed recovery: the new typed torn-record error
was correct at identity-read time but threw inside init-time recovery
walks, killing opens that previously survived. The boundary, redrawn:
- IDENTITY READS (get-by-id of a specific record, CAS blob point-get):
typed TornRecordError, unchanged — a caller who asked for THAT record
can act on the answer.
- SET-SHAPED READS AND WALKS (enumeration, pagination, batch hydration —
the paths recovery rebuilds and finds page over): HEAL PAST the torn
victim. The adapter's loud floor (error log + counted gauge) fires at
the encounter; the walk serves the remaining rows. One crash casualty
can no longer kill every query on its shard — or the open itself.
- WRITES OVER TORN RECORDS ARE THE CURE: the save path's read-merge, the
commit path's before-image capture, and the operations' rollback
captures all treat a torn prior as the create sentinel, narrated — the
incoming bytes replace the unreadable ones, and history for the id
honestly restarts at that generation. Corruption can never block its
own heal.
- THE NaN SOURCE: torn mapper state (nextId/entries carrying garbage)
discards with narration and re-derives via the existing rebuild path;
the mint gains a source guard healing a non-integer counter from the
live map. The reopen and first-write RangeError shapes are dead at the
source, both authority branches.
Pinned with the exact fault-injection scenarios: a torn entity record
(including the VFS root) no longer kills the open — walks heal past it,
the keeper rows serve, and the identity read of the victim itself is
typed-or-healed; a torn mapper reopens and mints sanely on the first
post-recovery write.
Gates: tsc 0 · unit 2065/2065 · integration 828 · conformance 31/31.
THE DEFAULT FLIP (ruled on proven evidence — at-ack survived 301/301
acked-writes-through-power-cut in block-layer fault injection; deferred
tree authority demonstrably loses flush-covered acks): a brain with NO
stored authority artifact now ADOPTS LOG AUTHORITY AT OPEN. The oracle
gates the flip exactly as the guarded adoption path always did — curable
divergences baseline-backfilled, the flip lands ONLY on a green verdict —
and a brain that cannot verify STAYS tree-authoritative loudly, with the
refusal recorded on the switch artifact so subsequent opens are cheap.
config logAuthority: 'defer' is the explicit documented opt-out (no
automatic adoption; declared flush-window loss; adoptLogAuthority() flips
later). A stored artifact always wins. RELEASES.md carries the posture.
Two standing .fails debt pins FLIP TO HOLDING under the default: the
at-ack crash-survival gap and the ack-at-log durability target — both now
permanent asserted truths, not aspirations.
POWER-CUT THROW SITES (fault-injection findings, brainy-alone config):
- A manifest-listed-but-unloadable column segment QUARANTINES at
discovery (loud once, counted always, quarantinedSegments() exposed for
the heal) and the field serves its remaining segments DEGRADED — never
a raw throw killing every query on the field. Real storage faults still
propagate untouched.
- Torn generation artifacts (NaN/garbage in manifest or counter) DISCARD
with narration at the store's open and recovery re-derives — plus a
defensive finite-integer guard at the init consumer. Never a RangeError
killing an open.
THE LOUD TORN-RECORD CONTRACT: an existing-but-unparseable stored record
now surfaces as a typed, counted TornRecordError on every entity-read
surface (including fifteen previously-blind per-item batch catches);
ENOENT stays clean-absent; artifact readers with designed absent-recovery
keep null-tolerance behind the loud floor. Disk corruption can no longer
read as silent data invisibility.
Suite migration: the default's pins inverted deliberately, generation
baselines made relative, quarantine-contract pins rewritten to the ruled
behavior.
Gates: tsc 0 · unit 2065/2065 (159 files) · integration 826 (93 files) ·
conformance 31/31 · kill-matrix 15/15 · torn-open guards 2/2.
An internal cross-engine fault-injection run (frozen-platter power-loss
capture) surfaced three release-gating findings; each cured in its owning
layer, each pinned:
1. WHOLE-LOG REPLAY ON UNCLEAN OPEN (the big one): log-authority replay
only covered facts ABOVE the manifest — but live canonical entity
writes are tmp+rename without per-file fsync, and the group-commit
flush syncs staging + manifest, never the live tree. Power loss could
therefore vaporize acked canonical bytes BELOW the manifest while the
log held every fact scan-clean (measured: 299 of 301 acks lost).
Now: a clean close stamps a clean-shutdown marker (fsynced, written
last); every open consumes it; an UNCLEAN open under log authority
folds the ENTIRE log into canonical — whole-entity after-images make
the re-apply idempotent and byte-safe. Zero cost on the happy path;
crash recovery pays one narrated fold. Recovery is replay: a crash is
just bigger lag.
2. TORN WRITER LOCK: power loss legally leaves the lock file present but
empty; the parse failure read as 'no holder' while the O_EXCL claim
EEXISTed forever — a PERMANENT lockout no staleness check could clear.
An unparseable lock is stale by definition (no live holder has one):
unlink loudly and re-loop; a racer rewriting a valid lock first wins.
3. PAIR GUARD: flush() called metadataIndex.stampWatermark unguarded;
a replacement metadata provider without the method killed the pair at
first flush. All three stamp calls are optional-chained — a missing
stamp is a verdict-side rescan, never a flush crash.
Pins: whole-log fold restores rows vanished below the manifest ·
clean-shutdown marker lifecycle (stamp/consume/re-stamp) · torn-lock
recovery with a fresh write after · stampless-provider flush.
Gates: unit 2055/2055 · integration 824 · kill-matrix 15/15.
An internal report from cross-engine write-path instrumentation: with
individual writes slower than the idle window (a contended disk), every
inter-write gap looked idle and fired a background full flush — 15 extra
flushes during 100 contended adds, amplifying the very pressure that
slowed the writes. The law now: an idle fire landing within the spacing
floor of the last flush DEFERS to the floor boundary instead of flushing;
the floor is min(interval, 10× the CONFIGURED idle window) — scaled to
caller intent (a tiny idle window keeps fast idle-driven durability;
default 2s/30s config gets a 20s floor), derived from the configured
idle, never from a deferred re-arm delay (which would compound into
runaway deferral). Deferred is never dropped: a lone write on a
then-quiet store still persists at the floor without any further write
arriving.
Pins: the contended-shape pin (six slow-spaced writes fire ≤2 idle
flushes, not one per gap; then still persist) + the original quiet-store
idle pin unchanged. Unit 2055/2055.
The generic reprojection engine (pure TS; the twin of the native
implementation — same frozen contract, one shared conformance intent):
register any ProjectionAdapter; advance(family, {budgetMs}) folds facts
from the adapter's own watermark to the head in installments ≤50ms with
real macrotask yields; foreground door traffic bumps the DoorSignal and
an in-flight advance yields within one installment ('preempted');
advanceAll round-robins families fairly. swap(family, buildAdapter) is
the doors-open migration primitive: the OLD projection keeps serving
while the new one builds beside it, the flip is atomic at parity, and a
concurrent second swap refuses typed. A fact the fold cannot apply
(typed ProjectionApplyError) is QUARANTINED — skipped, ledgered,
narrated per-doubling, exposed for refuse-affected-reads — the service
class law's fourth answer: never a wedged rebuild, never a silent skip.
The engine never writes stamps: each adapter owns its durability and its
stamp-after-data discipline. Upgrade, heal, and rebuild are now the same
machinery behind open doors.
FactLogSource wires any host's fact scan in one line
(factSourceFromHost(brain)); window-contract violations are loud.
Pins: 23 unit (budget resume without refold · preemption within one
installment · round-robin fairness under a skewed backlog · build-beside
visibility mid-swap · atomic flip · single-flight refusal · quarantine
skip/ledger/doubling · non-typed throw aborts · losing adapter
discarded) + 3 integration on a real brain (fold matches ground truth ·
doors answer mid-fold with the preemption path exercised · crash
mid-fold resumes from the stamp, never refolds).
Gates: unit 2054/2054 (157 files) · integration 820 (93 files) ·
conformance 31/31.
The private recovery discipline, applied to its own machinery: pending-
embed markers stop being sidecar files and become first-class log records
riding the write's OWN commit fact — embed.pending lands in the same
atomic append as its after-image (a marker can never be orphaned from its
write, or vice versa; in durable-at-ack mode it shares the write's
covering fsync — zero extra syncs), and the worker's landing commit rides
embed.landed with the inline vector. Crash recovery is now a FOLD of the
log (pending without a matching landed = recovered), skipped wholesale on
brains with no v2 history; the one-time legacy bridge folds existing
sidecar files in, migrates them as one fact, and deletes them —
idempotent under a crash mid-bridge. No code path writes the sidecar
again.
Plus the ENTITY-TRUTH digest law, found by this train's own pins:
canonical vector wrappers denormalize HNSW residue (connections + the
randomly-assigned node level) that the log deliberately does not carry —
the verification oracle digested it and would have reported false
state-differs on ~any nonzero-level node (a ~15% flake in the cutover pin
was the symptom). Both sides of every oracle comparison now normalize to
entity truth (nounEntityTruth); index residue has its own rebuild path
and is not entity state.
Pins: embed-markers-in-log 5/5 (same-generation marker, landed+fold-to-
zero, crash recovery via the log with the sidecar prefix EMPTY on disk,
legacy bridge, VFS hung-embedder ack) · deferred-embedding 5/5 unchanged
(the contract outlived its mechanism) · kill-matrix 11/11 · cutover 5/5
×10 runs (flake dead) · unit 2031/2031.
One deterministic v2 log (nine facts covering every fold-relevant behavior:
genesis, after-images with minted ints, a deferred embed pending→landed,
a sameAsGeneration vector ref, a verb, a tombstone, and an all-deduped
empty commit) whose ENCODED BYTES and FOLDED STATE are both pinned by
sha256 literals. The fixture (tests/fixtures/golden-log-v2.bin, 4128 B,
byte-verified against the encoder on every run) is the shared artifact a
second reader implementation consumes — it must reproduce the identical
fold digest; the pair is normative on disagreement. The fold law is
stated in prose beside the code: generation-ordered latest-per-id,
tombstone masking, embed.landed vector application, single-hop ref
resolution, key-sorted digest.
Also: decodeGroupV2 discriminated pad filler by RECORD COUNT, silently
swallowing legitimate empty commits (an all-deduped batch at a real
generation). Pads carry generation 0 — which writers can never mint — so
the generation is the honest discriminator; empty commits stay visible.
Pins: 4/4 (encode-exact, fixture-identical, fold-exact, human-readable
spot checks beside the hashes).
- Watermark stamping fans out at flush: all three projections stamped
with the committed generation before their flushes persist.
- waitForIndexed(path?, {generation, timeoutMs}) — the one honest read
barrier for write-then-recall consumers; typed timeout error carries
the pending count and names the gauge; getIndexStatus() gains
per-projection gauges. awaitPendingEmbeds() unchanged underneath.
- adoptLogAuthority() self-backfills curable divergences (pre-log
records, witness drift) by identity re-commit before flipping — a
fresh brain flips clean; log-ahead divergences still refuse loudly.
- The verification oracle gains VERB legs (all four divergence classes;
unwired = honest verbsChecked: 0, never a scope claim).
- find({where: {}}) match-all serves (was silent-empty, warm AND cold;
same fix in count/streaming/subgraph seeding); removeMany({where:{}})
refuses typed — a match-all bulk delete must be explicit.
- Aggregation native envelope stamped via noteSourceGeneration before
serializeState; the native-blob restore gates through the same
adoption verdict as caller-side state (the unconditional adopt dies).
- LC8 pinned: a wholesale directory move opens and serves identically
across all three intelligences, with history traveling.
Gates: unit 2031/2031 (156 files) · integration 812 (91 files) ·
conformance 27/27.
Every persisted projection artifact (metadata field indexes + column
segments, HNSW node records, graph adjacency LSM trees) now carries a
stamp asserting 'this state reflects every committed generation ≤ W,
atomically' — written LAST in each owner's flush (stamp-after-data: a
crash between data and stamp = unstamped = rescan, never trust). At load,
each owner computes the three-way verdict: stamped==committed → adopt
(zero work) · behind → catchup (gap reported) · above/unstamped → RESCAN,
loudly. Legacy artifacts re-derive once, then are stamped forever. Shared
law in projectionWatermark.ts (the aggregation verdict machinery,
generalized); vector artifacts carry model dimensions. Verdicts are
computed and exposed (watermark()/watermarkVerdict()/watermarkGap());
rebuild triggers unchanged — acting on 'catchup' is the fold train.
Pins: 22 unit (7 metadata · 8 hnsw · 7 graph, incl. spy-order
stamp-after-data) + the end-to-end reopen-adopts pin.
The cutover: new tail segments write format v2 (per-record [type, version,
cipherFlag, keyId] envelope; noun/verb after-images carry dense ints
MINTED AT APPEND from the id mapper — a rebuilt mapper reproduces
assignments exactly; log.genesis opens every new log with the id-space
width + a minted brain id; sync() seals to the header-declared sector
boundary with reader-invisible pad frames). Existing v1 segments are
never rewritten — per-segment decoder dispatch reads both formats and v2
facts map to the exact CommitFact shape all consumers already read.
Cutover on a live v1 log: an empty v1 tail re-heads in place; a non-empty
one is sealed by rotation, byte-identical. Records reserve the encryption
fields (cipherFlag 0 / keyId nil are the only legal values; anything else
refuses typed naming the needed newer reader) — crypto-ready with no
future bump on the compat surface. Empty-records facts are legal (an
all-deduped batch is a real generation — v1 semantics preserved; the
refusal there tore a column-store flush mid-commit in the full suite, the
consistency guard caught it loudly, and the root is fixed).
Golden byte vectors pinned for the second (native) reader implementation.
Pins: cutover 5/5 · codec 54 · kill-matrix stays 11/11.
The time-travel recall row moves from envelope-note to contracted: vector
search at a pinned past generation serves the vectors AS THEY STOOD —
a later re-embed never leaks into an earlier pin (byte-exact), tombstones
mask, the deferred-embed pin serves the stub on the vector leg until the
landing generation (text/metadata legs unaffected — triple intelligence by
design), and beyond-head pins refuse typed. Brainy-alone leg = the
documented ephemeral at-generation materialization; the at-scale leg rides
the accelerated provider's as-of index. Registry row added (shared ID
pending the master table).
Two release-blocking findings from the durability kill-matrix, both fixed
in the owning layer:
1. LOG-AUTHORITY REPLAY AT OPEN: durable-at-ack fsynced the fact before
the ack, but open() truncated every fact above the manifest — after a
power loss that takes the un-fsynced tmp+rename canonical bytes, the
acked write's ONLY durable copy was discarded. Now: under 'log'
authority, open() REPLAYS intact facts above the manifest into
canonical (FactLog.peekFactsAbove — CRC-gated, order-sorted) and
advances the manifest to cover them; tree-authority brains keep the
truncate contract they were promised. Pinned end to end: the power-loss
row constructs the exact disk state (fsynced log, vanished canonical
rename) and the acked write lives.
2. NO SILENT COMMIT: commitSingleOp buffered the generation BEFORE the
fact append; an append failure (ENOSPC) rejected the caller but the
next flush durably committed the generation with NO fact — a permanent
silent log gap. Now the failure path un-buffers and returns the counter
reservation: nothing commits, the log stays gap-free, and the canonical
execute-residue orphan is the documented crash-equivalent.
Plus: the kill-matrix itself (11 rows — every commit-path fault point ×
reopen-as-crash recovery contract, at-ack variants, disk-full row; five
new zero-cost faultPoint sites), the log-authority pin suite (oracle
green/red/state-differs, flip refusal, switch survives reopen, 9/9), and
the group-commit covering pins (5/5).
Gates: unit 2002/2002 (152 files) · integration 785 · conformance 27/27.
The provider contract (metadata addToIndex/removeFromIndex, vector
addItem/removeItem, id-mapper getOrAssign/remove) gains an optional
trailing generation — evaluated lazily at operation execute time (the
graph surface's thunk pattern, generalized), threaded from all 17
construction sites: undefined during generation-0 bootstrap, the real
committed generation everywhere else. Optional = additive: no existing
provider or caller breaks; native delta logs that stamped literal zero
start hearing truth. JS twins accept the parameter with parity notes.
Pins: provider doubles capture and assert nonzero monotonic generations
across add/update/remove on both surfaces.
The two-implementation contract surface as one pure module (no I/O):
segment header v2 (formatVersion 2 + sealSize in the reserved bytes),
per-record [type u8, version u8] envelope killing the unknown-kind
misclassification trap, the 12-type registry (after-images with minted
ints, tombstones, batch.meta, embed.pending/landed, blob.manifest,
projection.note, bootstrap.baseline, log.genesis with id-space width and
TYPED width-mismatch refusal), vectorLeg inline|{sameAsGeneration} with
writer-enforced single-hop, sector-sealed groups with pad frames, torn-tail
discipline, and GOLDEN BYTE VECTORS pinned so a second (native) reader
implementation can conform byte-for-byte. 50 format pins + a
fault-injecting storage wrapper (tear/drop-sync/fail-append) with 13
self-tests. v1 segments remain readable; nothing writes v2 yet — the
live-format cutover is its own commit.
A3 of the service-class pair (BRAINY-PROD-LATENCY-TRIAD): a VFS file write
ran the embedder synchronously while the caller waited — 5.6s p50 / 21.4s
p95 per small file on a production deployment, the dominant stage of every
capture write.
- add()/update() gain deferEmbedding: the write acks at durability (data +
metadata persisted, a DURABLE pending marker under
_system/pending_embeds/<id> written BEFORE the commit — orphan-safe
direction); the single-flight background worker embeds the CURRENT data
and swaps the vector in ATOMICALLY (ReplaceInVectorIndex — the row is
never absent from search; a deferred UPDATE keeps serving the OLD vector,
stale-beats-absent per the flicker law). Typed refusals: defer+vector,
defer-without-data.
- CRASH-SAFE: markers are recovered at open by a BOUNDED prefix listing
(never a store walk) and the worker resumes in the background — a crash
can delay a vector, never lose one. A wedged embedder trips a LOUD 60s
hang guard and the worker moves on (marker retained for retry).
- The honest gauges: getIndexStatus().pendingEmbeds + pendingEmbedCount();
awaitPendingEmbeds() is the eventual-vector-index BARRIER for callers
and tests that need searchability before proceeding.
- VFS adopts it everywhere a write path could wait on the embedder:
writeFile (both branches) and directory creation. Pinned in the
strongest form: writeFile resolves while the embedder HANGS FOREVER.
Pins: deferred-embedding 5/5 (ack law · stale-beats-absent · crash
recovery across sessions · VFS hung-embedder ack · typed refusals).
Gates: unit 1928/1928 · integration 765 · conformance 27/27.
DP6/DP8 of the Path Registry (BRAINY-PROD-LATENCY-TRIAD, the proven flicker
mechanism): update paths staged RemoveFromVectorIndex then AddToVectorIndex
as two separately-awaited transaction ops — between them a live row was in
NEITHER index (dark to semantic recall, fine in metadata list). The native
pair widened that window to seconds in production before their side's
visibility-commit fix; the structural cure lands here:
- hnswIndex.updateItem: absent → add; SAME vector → pure no-op (the
production shape — a type-only update re-indexed an unchanged vector,
remove+add did pure damage); changed vector → the node NEVER leaves the
index: synchronous vector swap first (every query from that instant sees
correct distances), then unlink/relink at the node's existing level via
shared internals (linkNode/unlinkNodeEdges refactored out of add/remove;
entry point and maxLevel provably unchanged).
- ReplaceInVectorIndexOperation: ONE transaction leg; feature-detects
provider updateItem (native seam flagged — their side ships updateItem,
then the adjacent remove+add fallback is dead code). Both update staging
sites swapped; delete sites untouched.
- LAZY-OPEN GATE (fleet adoption find, SELF-ENGINE-PAIR-STANDARD): under
disableAutoRebuild, ensureIndexesLoaded assessed ONLY the vector index —
a not-ready native METADATA provider never blocked the completion latch
and every find() silently returned [] on a populated store. All three
providers now vote; any not-ready report falls through to the rebuild.
- docs/path-registry.md: brainy's twin table for the 32 shared path IDs —
service class, budgets, lifecycle, narration, and the cited pin per row;
owed rows named (LC4 doors-open migration, MT4 yielding heals, LC7
downgrade contract) per the lifecycle-sprint choreography.
Pins: update-item-atomic 9/9 (visibility-atomic swap, reverse-index parity
vs fresh rebuild, entry-point invariants) · lazy-notready-honor 2/2.
Gates: unit 1928/1928 (148 files) · integration 760 · conformance 27/27.
A4 of the service-class pair (SELF-ENGINE-LIFECYCLE-SPRINT, David-directed:
'why do we need manual flushes at all?'). The production disease: 829
caller-scheduled per-write flushes convoying into 45-66s write walls —
cadence hand-rolled a layer above the only layer that can see dirty-node
counts and IO pressure.
- BrainyConfig.persistence: policy 'auto' (DEFAULT) | 'manual', with
flushEveryWrites (512) / flushIntervalMs (30s) / flushOnIdleMs (2s)
triggers. Auto = the engine kicks ONE single-flight BACKGROUND flush at
a threshold or when the store goes quiet; write acks NEVER await it (a
hung flush cannot block a write — pinned); a failed background flush is
LOUD and re-arms the trigger. 'manual' restores caller-owned cadence.
- Triggers wired at both write chokepoints (single-op post-commit +
transact post-commit); idle timer unref'd; close() tears the timer down
and drains the flight before its own final flush.
- RECOVERY SEMANTICS documented on the config: canonical records are
durable per-write regardless of policy — a crash between background
flushes loses derived state only, which converges at next open (epoch
machinery + the new incremental aggregation catch-up), bounded by the
un-flushed window. Never data loss.
Pins: write-count trigger fires one background flush with zero caller
calls · idle trigger · manual never self-flushes · THE ACK LAW (writes
acknowledge under a never-resolving flush). Gates: unit 1917/1917 ·
integration 760 · conformance 27/27 — green WITH auto as the default.
SELF-ENGINE-LIFECYCLE-SPRINT + BRAINY-PROD-LATENCY-TRIAD, the four asks:
(a) brain.flush() persists aggregation state stamped at the committed
generation. The stamp used to advance only at close(), so a long-lived
writer that flushes but never closes — the primary production shape —
left every write window behind the stamp, and ANY unclean exit forced a
whole-store backfill walk (per-entity work, measured >60s and
door-starving on a 9k-row production brain) on the first stats call.
(b) BEHIND-stamp adoption becomes adopt + INCREMENTAL CATCH-UP: the exact
missing window (stamp, committed] resolves its affected-id set from the
fact log and reconciles each entity with time-travel before/after reads
(asOf at both window bounds) through the same delta algebra the live
hooks use — cost bounded by writes since the last flush, never store
size, and exact under interleaving because reconciliation targets the
FIXED window end while later writes chain through hooks. Oversized
windows (>5000 affected) and unreadable windows demote to the announced
rescan — never a silent partial serve.
(c) The native provider's parallel rebuildAggregate — on the contract since
8.x but never invoked anywhere — is now the backfill walk's preferred
door: one call per aggregate with source-matched entities, replacing
the per-entity FFI stream.
(d) A delete whose before-image is unavailable can no longer SKIP the
aggregation hook silently (counts drifted upward forever): both delete
paths (remove() and transact) flag an exact rescan, loudly.
Pins: integration (flush stamp; unclean-exit reopen → exact counts through
an add + group-move + delete window with the walk spy proving ZERO
whole-store walks) + unit (provider rebuild invoked once with filtered
entities; flagAllForRescan; reconcile delta algebra). Gates: unit 1913/1913
· integration 760 · conformance 27/27.
BRAINY-PROD-LATENCY-TRIAD Track A1 (David-approved plan): the sort path's
value resolution goes BATCHED — one chunked metadata-record batch pass
serves any N, replacing the serial per-row getNoun loop (62-98ms x 3,224
rows = the measured 199-317 second silent scan on self prod). The
metadata record carries every sortable value: system scalars EXACT
(bucketed-index precision loss can never force a per-row disk read
again) and the user bag via the shape-aware split, both record eras.
- resolveOrderValuesBatch: the one sanctioned value source for ordered
reads (batch door: getNounMetadataBatch -> getMetadataBatch -> chunked
parallel; never serial).
- Column top-K page re-sort and the no-column fallback both rewired.
- B2 down-payment: the no-column fallback ANNOUNCES itself once per
field past 500 rows - silent degradation is illegal.
- THE CALL-SHAPE PIN (tests/unit/utils/metadataIndex-sort-callshape):
zero vector-record reads, batch calls only, latency-blind so it holds
on any machine - the serial loop cannot quietly return. Ordering
contract re-pinned through the batch path (nulls last both directions,
ties by id, never drop).
(! = perf contract change only; no API change. Gates: unit 1904/1904,
integration 758, conformance 27/27.)
The write side of the law, ruled 2026-08-03: data is either in main space
where developers can use anything, or it is in system.*.
- The reserved-name write door DIES: add/update/relate/updateRelation
metadata bags accept EVERY name (confidence, type, id, data, level,
content, ...) as ordinary user fields — indexed, filterable, sortable,
aggregatable, identical to any other field. The remap/enforce/warn
machinery, the reservedFieldPolicy config (now a typed init refusal),
and the compile-time metadata key bans are all removed. The one write
refusal left: keys spelled 'system.*' (namespace forgery), now enforced
on all four write doors.
- STORED RECORDS GO NESTED (v2): engine fields top-level, the user bag
nested verbatim under 'metadata', sealed by a format stamp — by-name
storage discrimination is unsound once colliders are admitted. Legacy
flat records stay readable forever through the shape-aware splitters
(sound for them: the old door refused colliders). Time travel rides the
same split (generation store snapshots whole records).
- Name-based index exclusions DIE: user frame indexes every name; the
excludeFields/indexedFields knobs and their silent-[] holes are gone;
bulk-payload protection is value-shape only, uniform across names.
- Consumer-sweep findings fixed in the same wave: per-type counts read
the frozen 'system.type' column (addToIndex sort, affinity tracking,
cold-count rehydration, VFS type bitmaps — legacy 'noun' fallback for
pre-rebuild reads); resolveHiddenIds addresses 'system.visibility'
(bare 'visibility' was a silent no-op under the law — VFS/system
entities leaked into default reads).
- Fidelity fallout fixed in the owning layers: readEntityFieldAddress
reads the bag first (colliders were absent-shadowed by its own guard)
and never serves system addresses from the bag; blob history refs read
the bag shape-aware; migration transforms now receive ONE normalized
view (engine fields + nested bag) regardless of stored era, and stray
flat-habit keys refuse with the fix in the message.
- THE REOPEN-COLLIDER CONFORMANCE CASE (required before any RC counts as
gates-green): all ten collider names + plumbing names written as user
fields, verified verbatim + queryable across live reads, flush+reopen,
a forced epoch rebuild, and asOf time travel; relation mirror; forgery
refusals; legacy flat-record compat. 8/8 green.
Gates: unit 1901/1901 (exit 0) · integration 758 (exit 0) · conformance
27/27 (exit 0) · consumer test sweep migrated (10 files).
Also completes the v8.10.2 write-granularity law for the transact() plan
path: a metadata-only batch update never rewrites the vector-bearing noun
record (planUpdate staged the unconditional save the update() fix removed).
Seven pins in tests/integration/level-field-shadow.test.ts including the
reporting consumer's exact repro rows; orderBy JSDoc documents the ordering
contract and the announced field-addressing law.
Also: idle PathResolver stats tick no longer logs NaN% every minute (logs
only on new traffic, via prodLog); graph-lsm-* key family recognized as
system resources (kills the per-boot unknown-key warning on provider-backed
brains). Four regression pins in tests/integration/update-write-granularity.
A production deployment's warm report showed metadata: 'unavailable' under a
native metadata provider. brain.warm()'s metadata leg only duck-typed the
built-in JS manager's hydrateAll() method, which a native provider has no
reason to implement.
- MetadataIndexProvider (src/plugin.ts) gains an optional warm?(): Promise<void>
hook, mirroring the existing vector and graph provider hooks. brain.warm()
now checks the active provider's own warm() FIRST, falls back to the JS
manager's hydrateAll() when absent, and reports 'unavailable' only when
neither exists -- never init() as a stand-in, since a native provider's
init() may be a cheap verify rather than a real warm.
- Tests (tests/unit/brainy/warm.test.ts): a live provider instance shaped to
have warm() reports 'warmed' and the hook called with no hydrateAll
fallback; shaped to have neither hook reports 'unavailable' (pins the
honest branch); the unmodified built-in JS manager still reports 'warmed'
via hydrateAll(), unchanged.
Additive scope agreed mid-flight with the native-provider team: a
maintenance-debt observability seam so an operator sees a grind coming
instead of discovering it as a CPU storm.
- New optional maintenanceDebt?(): Promise<ProviderMaintenanceDebt> hook on
all three provider contracts (vector, metadata, graph -- the same three
warm?() lives on). ProviderMaintenanceDebt is fields-all-optional: a
provider reports only what it truly measures (pendingBytes, pendingItems,
lastPassCompletedAt, lastPassOutcome, converging), never an estimate
dressed as fact.
- New public brain.maintenanceDebt(): a pure passthrough -- for each surface
it calls only the active provider's own hook and reports the payload
verbatim, or 'unavailable' when absent. No thresholds, no polling, no
JS-side estimation; the provider owns the numbers, the operator owns the
policy.
- ProviderMaintenanceDebt, MaintenanceDebtReport, and MaintenanceDebtOutcome
are exported from the package root.
- Tests (tests/unit/brainy/maintenance-debt.test.ts): hook present reports
'reported' with the exact payload passed through; hook absent reports
'unavailable' on every surface; mixed surfaces resolve independently of
each other.
RELEASES.md gains the 8.10.1 entry covering both fixes above and this
feature, including the no-hot-retry contract from the prior commit.
A production incident: a native-provider op ground 38-40s inside a transaction,
blew the apply-phase budget, rolled back, and a downstream pipeline hot-retried
the identical operation into a 6-minute CPU storm. Brainy itself never
auto-retried the timeout; the gap was that TransactionTimeoutError only said
"retryable" in prose, with nothing machine-readable for a caller to branch on.
- TransactionTimeoutError gains two typed, always-true fields: retryable
(a later attempt may succeed once the slowness resolves or the budget is
raised) and hotRetryUnsafe (an immediate identical retry re-pays the full
cost that just timed out and can cascade into a CPU storm -- callers must
latch and back off, never loop). context's existing telemetry fields
(timeoutMs, operationIndex, elapsedMs, totalOperations, operationName) are
now documented as the caller's backoff inputs.
- Updated the "retryable" doc-prose sites (transact()'s timeoutMs option,
transactionBudgetFloorMs, Transaction.execute()'s contract) to point at
the new fields instead of bare prose.
- Regression pin (tests/unit/transaction/timeout-never-internally-retried.test.ts):
an execution counter proves the engine never re-drives a timed-out
operation, through both the single-op engine TransactionManager/Transaction
drives for every single-record write, and add()'s upsert-race retry loop
(which must exit on the first TransactionTimeoutError, never treat it like
the lost-insert-race signal it retries on).
- Removed TransactionManager.executeTransactionWithResult -- zero callers
anywhere in the codebase.
Reconciles the vector-index rename to the ruled three-layer naming: the
provider contract's identity field is now a REQUIRED readonly name (was
optional providerId), self-reported and truthful, rendered as
[vector-index:<name>] where the index identifies itself and stamped into
the op-name strings journals already parse (AddToVectorIndex(<name>)).
The built-in JS engine names itself hnsw-js. A runtime provider instance
compiled against the previous optional contract is tolerated - never
crashed on, never silently mislabeled: it stamps unknown-provider and
emits one loud warning naming the missing field. Graph index operations
keep their static names (they never interpolate provider identity), and
no public API exports a provider-routed hnsw-carrying name, so no
deprecation shim is required.
Three pieces addressing the cold-restart-write incident where a production
deployment's first writes after every restart (33-35s each on a cold page
cache) blew the op-count-scaled transact budget mid-batch: every write is
itself a multi-op transaction, so one cold operation consumed the whole
budget, the gate before the next operation tripped, and the write rolled
back atomically - refused, retried, and refused again until the page cache
warmed passively.
- The budget's start-gating contract is now explicit and pinned: it gates
STARTING the next operation, never rolling back completed work for
elapsed time (the shipped schedule since 8.7.0, now stated in contract
JSDoc, guarded by code for operation 0, and enforced by regression
tests). The 30s floor is configurable via transactionBudgetFloorMs for
stores whose cold operations legitimately run long.
- New brain.warm() eagerly loads the vector index, metadata index, and
graph adjacency so first operations after a cold restart run at
steady-state cost. Returns a WarmReport with an honest per-surface
outcome (warmed / probed / unavailable) - never reports a probe as a
warm. warmOnOpen: true runs it during init(). New optional provider hook
warm() on the vector and graph plugin contracts.
- Vector-index transaction op classes renamed from the backend-specific
AddToHNSWOperation / RemoveFromHNSWOperation to backend-neutral
AddToVectorIndexOperation / RemoveFromVectorIndexOperation, stamping the
active backend into the emitted op-name string (AddToVectorIndex(js-hnsw)
vs a native provider's own identity) so journals never misdirect an
operator toward an index that isn't running.
